README.md

---
license: apache-2.0
language:
- vi
tags:
- audio
- text
- speech-recognition
- music
- lyrics-alignment
- karaoke
- word-timestamps
pretty_name: Vietnamese Song Lyrics and Word Timestamps Dataset (Full Songs)
---

# 🎵 Vietnamese Song Lyrics and Word Timestamps Dataset 

## Dataset Summary
The `song_dataset` provides high-quality Vietnamese song data, including metadata, full lyrics, and particularly **word-level timestamps**.

This dataset is optimally designed for tasks such as:
- Training and evaluating automatic speech recognition (ASR) models on music.
- Lyrics synchronization (Lyrics Alignment / Karaoke generation).
- Natural language processing (NLP) analysis on song lyrics.

The data is stored in `JSONL` format and is split into two sets:
- **Train split:** `train.jsonl` (approx 230.62 hours of audio)
- **Test split:** `eval.jsonl` (approx 20.59 hours of audio)

## Data Structure

Each line in the `.jsonl` file is a JSON object representing a full song, which includes the following fields:

- `id` (string): A unique identifier for the song.
- `title` (string): The title of the song.
- `artist` (string): The performing artist(s).
- `album` (string): The album containing the song.
- `streaming_url` (string): A URL to extract the audio stream. *Note: These URLs may have a temporary lifespan (token expiration).*
- `lyrics` (string): The complete lyrics of the song, where sentences are separated by a newline character (`\n`).
- `word_timestamps` (list of lists of dictionaries): A 2D array representing the timing information for each word. The outer list represents sentences, the inner list represents words within the sentence. Each word object has:
  - `start` (int): The start time of the word (in milliseconds).
  - `end` (int): The end time of the word (in milliseconds).
  - `word` (string): The text content of that word.

### Example Data Instance

```json
{
  "id": "ZZ8CC7AZ",
  "title": "Đợi Chờ Bóng Xuân",
  "artist": "Eric Toàn Nguyễn",
  "album": "Trách Duyên Bẽ Bàng",
  "streaming_url": "https://...",
  "lyrics": "Lặng nhìn sương rơi\nMấy mùa mưa qua...",
  "word_timestamps": [
    [
      {"start": 39340, "end": 39850, "word": "Lặng"},
      {"start": 39850, "end": 40340, "word": "nhìn"},
      {"start": 40340, "end": 40850, "word": "sương"},
      {"start": 40850, "end": 43850, "word": "rơi"}
    ],
    [
      {"start": 43860, "end": 44370, "word": "Mấy"},
      {"start": 44370, "end": 44860, "word": "mùa"},
      {"start": 44860, "end": 45870, "word": "mưa"},
      {"start": 45870, "end": 48390, "word": "qua"}
    ]
  ]
}
```

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("<your-hub-username>/<dataset-name>")

# Access the first song in the training set
print(dataset['train'][0]['title'])
print(dataset['train'][0]['lyrics'])
```